Presenter: Devendra Tagare - DataTorrent Engineer, Contributor to Apex, Data Architect experienced in building high scalability big data platforms.
Apache Apex is a next generation native Hadoop big data platform. This talk will cover details about how it can be used as a powerful and versatile platform for big data.
Apache Apex is a native Hadoop data-in-motion platform. We will discuss architectural differences between Apache Apex features with Spark Streaming. We will discuss how these differences effect use cases like ingestion, fast real-time analytics, data movement, ETL, fast batch, very low latency SLA, high throughput and large scale ingestion.
We will cover fault tolerance, low latency, connectors to sources/destinations, smart partitioning, processing guarantees, computation and scheduling model, state management and dynamic changes. We will also discuss how these features affect time to market and total cost of ownership.
2. What is Apex
2
• Platform and runtime engine that enables development of
scalable and fault-tolerant distributed applications
• Hadoop native
• Process streaming or batch big data
• High throughput and low latency
• Library of commonly needed business logic
• Write any custom business logic in your application
3. Applications on Apex
3
• Distributed processing
• Application logic broken into components called operators that run in a distributed fashion across
your cluster
• Scalable
• Operators can be scaled up or down at runtime according to the load and SLA
• Fault tolerant
• Automatically recover from node outages without having to reprocess from beginning
• State is preserved
• Long running applications
• Operators
• Use library to build applications quickly
• Write your own in Java using the API
• Operational insight – DataTorrent RTS
• See how each operator is performing and even record data
7. Application Development Model
7
A Stream is a sequence of data tuples
A typical Operator takes one or more input streams, performs computations & emits one or more output streams
• Each Operator is YOUR custom business logic in java, or built-in operator from our open source library
• Operator has many instances that run in parallel and each instance is single-threaded
Directed Acyclic Graph (DAG) is made up of operators and streams
Directed Acyclic Graph (DAG)
Output
StreamTuple Tuple er
Operator
er
Operator
er
Operator
er
Operator
er
Operator
er
Operator
8. Advanced Windowing Support
8
Application window
Sliding window and tumbling window
Checkpoint window
No artificial latency
12. Partitioning and unification
12
NxM PartitionsUnifier
0 1 2 3
Logical DAG
0 1 2
1
1 Unifier
1
20
Logical Diagram
Physical Diagram with operator 1 with 3 partitions
0
Unifier
1a
1b
1c
2a
2b
Unifier 3
Physical DAG with (1a, 1b, 1c) and (2a, 2b): No bottleneck
Unifier
Unifier0
1a
1b
1c
2a
2b
Unifier 3
Physical DAG with (1a, 1b, 1c) and (2a, 2b): Bottleneck on intermediate Unifier
13. Advanced Partitioning
13
0
1a
1b
2 3 4Unifier
Physical DAG
0 4
3a2a1a
1b 2b 3b
Unifier
Physical DAG with Parallel Partition
Parallel Partition
Container
uopr
uopr1
uopr2
uopr3
uopr4
uopr1
uopr2
uopr3
uopr4
dopr
dopr
doprunifier
unifier
unifier
unifier
Container
Container
NICNIC
NICNIC
NIC
Container
NIC
Logical Plan
Execution Plan, for N = 4; M = 1
Execution Plan, for N = 4; M = 1, K = 2 with cascading unifiers
Cascading Unifiers
0 1 2 3 4
Logical DAG
14. Dynamic Partitioning
14
• Partitioning change while application is running
ᵒ Change number of partitions at runtime based on stats
ᵒ Determine initial number of partitions dynamically
• Kafka operators scale according to number of kafka partitions
ᵒ Supports re-distribution of state when number of partitions change
ᵒ API for custom scaler or partitioner
2b
2c
3
2a
2d
1b
1a1a 2a
1b 2b
3
1a 2b
1b 2c 3b
2a
2d
3a
Unifiers not shown
15. How tuples are partitioned
15
• Tuple hashcode and mask used to determine destination partition
ᵒ Mask picks the last n bits of the hashcode of the tuple
ᵒ hashcode method can be overridden
• StreamCodec can be used to specify custom hashcode for tuples
ᵒ Can also be used for specifying custom serialization
tuple: {
Name,
24204842, San
Jose
}
Hashcode:
001010100010
101
Mask
(0x11)
Partition
00 1
01 2
10 3
11 4
16. Custom partitioning
16
• Custom distribution of tuples
ᵒ E.g.. Broadcast
tuple:{
Name,
24204842, San
Jose
}
Hashcode:
001010100010
101
Mask
(0x00)
Partition
00 1
00 2
00 3
00 4
17. Fault Tolerance
17
• Operator state is checkpointed to a persistent store
ᵒ Automatically performed by engine, no additional work needed by operator
ᵒ In case of failure operators are restarted from checkpoint state
ᵒ Frequency configurable per operator
ᵒ Asynchronous and distributed by default
ᵒ Default store is HDFS
• Automatic detection and recovery of failed operators
ᵒ Heartbeat mechanism
• Buffering mechanism to ensure replay of data from recovered point so
that there is no loss of data
• Application master state checkpointed
18. • In-memory PubSub
• Stores results emitted by operator until committed
• Handles backpressure / spillover to local disk
• Ordering, idempotency
Operator
1
Container 1
Buffer
Server
Node 1
Operator
2
Container 2
Node 2
Buffer Server
18
20. Processing Guarantees - Recovery
20
Atleast once
• On recovery data will be replayed from a previous checkpoint
ᵒ Messages will not be lost
ᵒ Default mechanism and is suitable for most applications
• Can be used in conjunction with following mechanisms to achieve
exactly-once behavior in fault recovery scenarios
ᵒ Transactions with meta information, Rewinding output, Feedback from external
entity, Idempotent operations
Atmost once
• On recovery the latest data is made available to operator
ᵒ Useful in use cases where some data loss is acceptable and latest data is
sufficient
Exactly once
• At least once + state recovery + operator logic to achieve end-to-end
exactly once
21. Stream Locality
21
• By default operators are deployed in containers (processes) randomly
on different nodes across the Hadoop cluster
• Custom locality for streams
ᵒ Rack local: Data does not traverse network switches
ᵒ Node local: Data is passed via loopback interface and frees up network
bandwidth
ᵒ Container local: Messages are passed via in memory queues between
operators and does not require serialization
ᵒ Thread local: Messages are passed between operators in a same thread
equivalent to calling a subsequent function on the message
22. Next Gen Stream Data Processing
• Data from variety of sources (IoT, Kafka, files, social media etc.)
• Unbounded, continuous data streams
ᵒ Batch can be processed as stream (but a stream is not a batch)
• (In-memory) Processing with temporal boundaries (windows)
• Stateful operations: Aggregation, Rules, … -> Analytics
• Results stored to variety of sinks or destinations
ᵒ Streaming application can also serve data with very low latency
22
Browser
Web Server
Kafka Input
(logs)
Decompress,
Parse, Filter
Dimensions
Aggregate Kafka
Logs
Kafka
23. Batch vs. Streaming
Credit: Gyula Fóra & Márton Balassi: Large-Scale Stream Processing in the Hadoop Ecosystem
23
24. Architecture and Features
Spark Streaming Apex
Model micro-batch native streaming/data-in-motion
Language Java, Scala, client bindings Java (Scala)
API declarative compositional (DAG),
declarative*
Locality data locality advanced processing locality
Latency high very low (millis)
Throughput very high very high
Scalability scheduler limit horizontal
Partitioning standard advanced (parallel pipes,
unifiers)
Connector Library Limited (certification), Rich library of connectors and24
25. Operability
Spark Streaming Apex
State Management RDD, user code checkpointing
Recovery RDD lineage incremental (buffer server)
Processing Sem. exactly-once* end-to-end exactly-once
Backpressure user configuration Automatic (buffer server
memory + disk)
Elasticity yes w/ limited control yes w/ full user control
Dynamic topology no yes
Security Kerberos Kerberos, RBAC*, LDAP*
Multi-tennancy depends on cluster
manager
YARN, full isolation
DevOps tools basic REST API, DataTorrent RTS25
31. Maximize Revenue w/ real-time insights
31
PubMatic is the leading marketing automation software company for publishers. Through real-time analytics,
yield management, and workflow automation, PubMatic enables publishers to make smarter inventory
decisions and improve revenue performance
Business Need Apex based Solution Client Outcome
• Ingest and analyze high volume clicks &
views in real-time to help customers
improve revenue
- 200K events/second data flow
• Report critical metrics for campaign
monetization from auction and client logs
- 22 TB/day data generated
• Handle ever increasing traffic with
efficient resource utilization
• Always-on ad network
• DataTorrent Enterprise platform,
powered by Apache Apex
• In-memory stream processing
• Comprehensive library of pre-built
operators including connectors
• Built-in fault tolerance
• Dynamically scalable
• Management UI & Data Visualization
console
• Helps PubMatic deliver ad performance
insights to publishers and advertisers in
real-time instead of 5+ hours
• Helps Publishers visualize campaign
performance and adjust ad inventory in
real-time to maximize their revenue
• Enables PubMatic reduce OPEX with
efficient compute resource utilization
• Built-in fault tolerance ensures customers
can always access ad network
32. Industrial IoT applications
32
GE is dedicated to providing advanced IoT analytics solutions to thousands of customers who are using their
devices and sensors across different verticals. GE has built a sophisticated analytics platform, Predix, to help its
customers develop and execute Industrial IoT applications and gain real-time insights as well as actions.
Business Need Apex based Solution Client Outcome
• Ingest and analyze high-volume, high speed
data from thousands of devices, sensors per
customer in real-time without data loss
• Predictive analytics to reduce costly
maintenance and improve customer service
• Unified monitoring of all connected sensors
and devices to minimize disruptions
• Fast application development cycle
• High scalability to meet changing business
and application workloads
• Ingestion application using DataTorrent
Enterprise platform
• Powered by Apache Apex
• In-memory stream processing
• Built-in fault tolerance
• Dynamic scalability
• Comprehensive library of pre-built operators
• Management UI console
• Helps GE improve performance and lower
cost by enabling real-time Big Data analytics
• Helps GE detect possible failures and
minimize unplanned downtimes with
centralized management & monitoring of
devices
• Enables faster innovation with short
application development cycle
• No data loss and 24x7 availability of
applications
• Helps GE adjust to scalability needs with
auto-scaling
33. We Are Hiring
33
• jobs@datatorrent.com
• Developers/Architects
• QA Automation Developers
• Information Developers
• Build and Release
• Community Leaders
Partitioning & Scaling built-in
Operators can be dynamically scaled
Throughput, latency or any custom logic
Streams can be split in flexible ways
Tuple hashcode, tuple field or custom logic
Parallel partitioning for parallel pipelines
MxN partitioning for generic pipelines
Unifier concept for merging results from partitions
Helps in handling skew imbalance
Advanced Windowing support
Application window configurable per operator
Sliding window and tumbling window support
Checkpoint window control for fault recovery
Windowing does not introduce artificial latency
Stateful fault tolerance out of the box
Operators recover automatically from a precise point before failure
At least once
At most once
Exactly once at window boundaries